The driver arrives and nobody is home. Walk me through what happens next and what your system has to get right.

An attempt is an event with an outcome, not a success or a failure

The framing that produces a weak answer is that delivery either works or does not. The framing that produces a strong one is that every stop generates an attempt event, the attempt has exactly one outcome drawn from a defined set, and almost all of the difficulty in last-mile systems lives in that set being too small.

Consider what the driver can legitimately do at a door where nobody answers. Leave the parcel in a location the customer nominated. Leave it with a neighbour. Take it to a nearby shop or locker. Card the address and take it back to the depot. Refuse to leave it because the item requires a signature or an age check. Discover the address does not exist, or that the recipient has moved. Each of those is a different downstream flow, a different customer message, a different cost, and a different liability position, and a system whose event model has "delivered" and "failed" will collapse them into a status that tells nobody what to do next.

So the first thing to get right is the vocabulary, and the vocabulary has to be shared with whoever is upstream. If you are the retailer and a carrier reports its own codes, mapping those codes into your set is where the meaning is either preserved or lost, and an unmapped code defaulting to a generic exception is how a parcel sitting in a shop three streets away gets reported to the customer as delayed.

The outcome the driver may choose was decided before the van left

The most common design error is treating authority to leave as a decision made at the doorstep. It is not. It is the resolution of several inputs that were all knowable earlier, and the doorstep is simply where the resolution is applied.

The product decides some of it. Age-restricted goods require a check that only the recipient can satisfy, high-value items may require a signature, prescription items and anything with a controlled chain of custody cannot be left. The consignor decides some of it: a retailer may forbid neighbour delivery for its entire account, or require photographic proof on every parcel. The customer decides some of it, through a standing safe-place instruction or a preference expressed after despatch. And the address decides some of it, because a flat with a communal entrance has no safe place, whatever anybody has nominated.

Resolving all of that into a single instruction that travels with the parcel to the handheld means the driver's app can present only the outcomes that are permitted, which is both faster and considerably safer than relying on a driver to remember which account forbids what. It also gives you something to point at afterwards. When a customer disputes a doorstep delivery, the question is whether leaving it was authorised, and the answer has to be a recorded permission rather than a recollection.

The interesting complication is that the customer can change their mind while the parcel is on the van. Redirection to a neighbour, to a shop, or to a different day is a genuinely useful feature and it means the instruction is mutable up to a point of no return, which needs defining: once the round is loaded, once the van has left, or once the driver is within a stop or two. Accepting a change you cannot honour is worse than refusing it, because the customer now believes something that the network will contradict.

The parcel acquires its own lifecycle at the depot

A carded parcel goes back to the depot and becomes an object with a state, a location and a clock, and this is where systems that modelled only the outward journey run out of model.

stateDiagram-v2
  [*] --> OnVan
  OnVan --> Delivered
  OnVan --> LeftSafePlace
  OnVan --> HandedToNeighbour
  OnVan --> AtDepotAfterAttempt
  AtDepotAfterAttempt --> OnVan
  AtDepotAfterAttempt --> AtCollectionPoint
  AtCollectionPoint --> Delivered
  AtCollectionPoint --> ReturnToSender
  AtDepotAfterAttempt --> ReturnToSender

The transition worth examining is the loop from the depot back onto a van, because it is the one with a budget. Most operations allow a fixed number of attempts, and every re-entry consumes one and consumes a slot on a round that was planned without it. The parcel also needs a physical home in the depot with a recorded location, because a parcel that exists in the system as awaiting redelivery and cannot be found on the shelf is functionally lost, and the volume of those grows quietly.

The clock is the other half. A parcel held at a collection point has a holding period after which it is returned, and a parcel that has exhausted its attempts becomes a return to the consignor, which triggers a refund decision, a restocking movement and sometimes a customs consequence if a border was crossed. Returns are a whole flow that costs real money and is routinely designed last.

Proof of delivery is evidence for an argument that has not started yet

The purpose of proof of delivery is not to close the stop in the app, it is to settle a dispute weeks later with someone who is annoyed and possibly right. That changes what you capture and how you keep it.

A photograph is now the common artefact and it is useful in proportion to what surrounds it. On its own it shows a parcel on a doorstep, which could be any doorstep. With the location the device reported at the time, the timestamp, the scanned barcode of the parcel, the round and the driver, it becomes a coherent account of a specific event. Where a signature is captured, the name of the person signing matters as much as the mark, because "signed for" by an unnamed party is thin evidence. And where a neighbour accepted the parcel, the neighbour's address is the fact that lets somebody go and retrieve it.

The device is often out of signal at the moment of capture, so all of this queues locally and uploads later, which means the timestamp you trust is the one the device recorded rather than the one the server received. Getting that ordering right is the difference between a defensible record and a set of events that appear to have happened in the wrong sequence. Retention matters too: the evidence has to outlive the window in which a customer can claim non-receipt and a consignor can raise a chargeback, and those windows are usually longer than an engineer's instinct for how long to keep photographs.

The second attempt costs more than the delivery ever earned

The economics are what should drive the design, and they are simple enough to state in the interview. The marginal cost of a stop is the driver's time and the vehicle's distance, and a failed attempt incurs almost all of that cost while producing no delivery. A second attempt incurs it again, plus depot handling twice, plus the customer contact. On a typical parcel the margin does not survive two attempts, which is why first-time delivery rate is the metric the operation actually manages.

That reframes several decisions. Nominating a safe place at checkout is worth more than any routing improvement, because it converts a probable failure into a certain success. Diverting to a collection point on the first failure rather than the second removes a whole attempt from the network. Narrow, confirmed time windows raise the chance of someone being in, at the cost of route efficiency, and whether that trade is worth making depends on the local failure rate rather than on principle. And a genuinely useful pre-delivery notification, sent early enough that the customer can act on it, is one of the cheapest interventions available.

The failure to watch for is optimising the plan without measuring the outcome. A round that looks efficient on paper and produces a high card rate is more expensive than a longer round that delivers everything, and only per-stop outcome data tells you which you have.

Drivers under pressure, and the signals that reveal it

The last thing an experienced answer raises is that the attempt outcome is self-reported by a person whose day is measured on completion. Under time pressure, and particularly near the end of a round, outcomes get recorded that did not happen: a parcel marked as carded without an attempt, or marked delivered to a safe place that does not exist. This is not a reason to distrust drivers, it is a reason to design the record so that reality is easier to record than fiction and so that patterns are visible.

The signals are behavioural rather than technical. Attempt events whose reported location does not match the delivery address. Stops closed implausibly quickly after the previous one. A specific round or driver with a card rate far from its peers. Safe-place deliveries at addresses where no safe place was ever nominated. Photographs that repeat. None of those is proof on its own and each is a reason to look, and having them surfaced is what stops a customer complaint being the first time anyone notices.

The design centre of last mile is the outcome set of an attempt. Make it rich enough to describe what really happens at a door, decide upstream which outcomes are permitted, and the notifications, the depot flows and the disputes all become tractable.